MLM versus OLS

[Adrienne Wold]

Hello,


I am interested in estimating the following the dependent variable: Food security as a survey response to the question how often did you go without food in the last year (never, sometimes, frequently, always). I have 53935 individuals in my dataset from 36 countries, and they have been placed into 50x50km "Grids" (1824 unique grids in the data). My main explanatory variable of interest is the number of riots that occurred in the preceding year at the grid level, as well as the number of riots that occurred within the country overall. I have a host of controls at both the grid level (population, harvest area, nearest city distance, etc.) and some variables at the individual level from the survey ( gender, number of household members).

My question pertains to model fitting. I'm not sure whether the right approach would simply be to use OLS and aggregate the individual responses to the grid level to obtain a mean score for the food security variable and have the grid be the main unit of analysis and include region FE (larger units than grid) , or whether to use MLM (mixed command) and measure food security at the individual level with three levels: individual, grid, and country.


Thank you.

[Clyde Schechter]

It depends on your research goal. If you are attempting to describe the effect of riots on mean food security in each grid, then aggregating to grid level and using OLS is appropriate. That analysis will give you no insight into individual variation of food security within a grid, and it will markedly blunt your ability to adjust for, or estimate, the effects of the individual level variables (gender, number of household members). But if within-grid variation in food-security is of no importance to your goals and you do not care much about effects of gender and household members, then there is no reason to a more complicated MLM to get those things.

But if those things are important, you need an MLM to get them.